SMARTHEP Edge Machine Learning School

23 Sept 2024, 09:00 → 27 Sept 2024, 18:00 Europe/Zurich

222/R-001 (CERN)

Anna Sfyrla (Universite de Geneve (CH)), Maurizio Pierini (CERN), Sioni Paris Summers (CERN), Thea Aarrestad (ETH Zurich (CH))

Welcome to the NextGen and SMARTHEP School on Edge Machine Learning! 

This  training program is designed to provide participants with knowledge and hands-on experience in the emerging field of Edge Machine Learning, tailored for applications at the LHC.

The program includes introductory lectures, tutorials, and seminars covering a variety of related topics, with ample time for discussion. Tutorials include:

• Model compression techniques
• Efficient GPU inference
• Efficient FPGA inference
• Neuromorphic computing and spiking neural networks
  
• Heterogeneous accelerated computing

 

We will also feature talks on real-time inference applications for autonomous vehicles, satellites, gravitational wave physics, high energy particle physics and more.

Additionally, the program features contributed flash talks and poster sessions, allowing attendees to share their work and insights. And a PhD student poster prize awaits!

Monday 23 September

09:00 → 10:45 Lectures: ML

Convener: Anna Sfyrla (Universite de Geneve (CH))

09:00 Introduction to ML

Speaker: Michael Kagan (SLAC National Accelerator Laboratory (US))

Kagan_SMARTHEP_2024.pdf

09:30 Discussion

09:45 ML on the Edge at the LHC experiments

Speaker: Dylan Sheldon Rankin (University of Pennsylvania (US))

smarthep_lhcfastml_dsr_23sep24.pdf

10:30 Discussion

10:45 → 11:15 Coffee break

11:15 → 12:00 Lectures: GPUs

11:15 Introduction to GPUs

Speaker: NVIDIA team

GPUsIntro.pdf

11:45 Discussion

12:00 → 13:30 Lunch break

13:30 → 14:30 Lectures: GPUs

Convener: Thea Aarrestad (ETH Zurich (CH))

13:30 ML inference on GPUs

Speaker: NVIDIA team

MLGPUs.pdf

14:15 Discussion

14:30 → 16:00 Tutorial: GPUs

Convener: Thea Aarrestad (ETH Zurich (CH))

14:30 Tutorial on network optimization for GPUs

Speaker: NVIDIA team

GPUsTutorial.pdf

16:00 → 16:30 Coffee break

Coffee offered at the mezzanine outside the main auditorium

16:30 → 18:00 Flash talks / poster session: Flash talks and poster session in Auditorium Mezzanine

All contributors will present a poster, which they will need to print by themselves. Poster holders will be available on Monday at the mezzanine. Please also upload the pdf of the poster in the agenda.

Some contributors have also asked for a flash talk. The duration of the flash talks is 3 mins. Please upload the flash talk contribution by Monday morning 11am. Thanks!

Convener: Thea Aarrestad (ETH Zurich (CH))

all_flash_EDGEML.pdf

16:30 FlashSim: Towards a Digital Twin of the CMS Experiment using Normalizing Flows and Flow Matching

The simulation of high-energy physics collision events is a key element for data analysis at present and future particle accelerators. The comparison of simulation predictions to data allows us to look for rare deviations that can be due to new phenomena not previously observed. The CMS Collaboration is investigating how novel machine learning algorithms, specifically Normalizing Flows and Flow Matching, can be used to perform accurate simulations with several orders of magnitude of speed-up compared to traditional approaches, contributing to the development of a "Digital Twin" of the CMS Experiment, a simulation framework named FlashSim. The classical simulation chain computes energy deposits, electronics response, and reconstruction from a physics process. We propose an end-to-end approach, directly simulating the final high-level format from physical inputs, skipping intermediate steps. The speed and accuracy of the proposed approach make it a compelling tool for the present and future needs of CMS Collaboration.

Speaker: Francesco Vaselli (Scuola Normale Superiore & INFN Pisa (IT))

smarthep_flashtalk_vaselli.pdf

16:33 Improving the Inference of Graph Neural Networks for Track Reconstruction

Optimizing the inference of Graph Neural Networks (GNNs) for track finding is crucial for enhancing the computing performance of particle collision event reconstruction. Track finding involves identifying and reconstructing the paths of particles from complex, noisy detector data. By leveraging GNNs, we can model the relationships between detector hits as a graph, where nodes represent hits and edges represent potential connections between them. To speed up the inference of these GNN models, it is important to reduce computational overhead, improve model architecture, and exploit hardware accelerators such as GPUs. Techniques like quantization and pruning can be employed to minimize model size and inference time without sacrificing accuracy.

Speakers: Alina Lazar (Youngstown State University (US)), Henry Paschke, James jsgaboriaultwhit@student.ysu.edu, Jay Chan (Lawrence Berkeley National Lab. (US)), Minh Tuan Dang (CERN), Paolo Calafiura (Lawrence Berkeley National Lab. (US)), Xiangyang Ju (Lawrence Berkeley National Lab. (US))

ImprovingGNNInference.pdf

16:36 An Open-Source RISC-V-based GPGPU Accelerator for Machine Learning-based Edge Computing Applications

In recent years, the demand for real-time machine learning (ML)-based computing solutions has driven the rapid growth of edge computing. The adopted hardware must strike a delicate balance by providing sufficient computational power to meet stringent real-time constraints while minimizing energy consumption. General-purpose graphics processing units (GPGPU) are a commonly adopted solution to maximize the data parallelism of ML algorithms thanks to more specialized hardware.

This work presents an open-source RISC-V-based GPGPU accelerator designed to support research in the ML-based edge computing domain. The accelerator features a low-power GPGPU streaming multiprocessor (SM) and offers two selectable memory hierarchies: cache-based and scratchpad-based. Its high configurability, regarding the number of threads, warps, and memory sizes, enables matching specific application requirements. The accelerator has been integrated into our in-house-designed eXtendible Heterogeneous Energy Efficient Platform (X-HEEP) microcontroller to improve its data processing capabilities and provide a real-world integration example.

Speaker: Simone Machetti

poster.pdf

slides.pdf

16:40 Accelerating Machine Learning algorithms in FPGAs for the trigger system of a SiPM-based upgraded camera of the CTA Large-Sized Telescopes

Current Imaging Atmospheric Cherenkov Telescopes use combined analog and digital electronics for their trigger systems, implementing simple but fast algorithms. Such trigger techniques are used due to high data rates and strict timing requirements. In recent years, in the context of a possible upgraded camera for the Large-Sized Telescopes (LSTs) of the Cherenkov Telescope Array (CTA) based on Silicon PhotoMultipliers, a new fully digital trigger system incorporating Machine Learning (ML) algorithms is being developed. The main concept is to implement those algorithms in FPGAs to increase the sensitivity and efficiency of the real-time decision making while being able to fulfill timing constraints. The project is full of challenges, such as complex printed circuit board design, complex FPGA logic design, and translating high level ML models to FPGA synthesizable code. We are currently developing a test bench as a proof of concept and to evaluate the FPGA performance of the algorithms.

Speakers: Alejandro Pérez Aguilera (IPARCOS-UCM), Prof. Juan Abel Barrio (IPARCOS-UCM), Dr Luis Ángel Tejedor (IPARCOS-UCM)

FlashTalk_APerezAguilera.pdf

PosterAPerezAguilera.pdf

16:46 Nanosecond ML for calorimeter segmentation

Effective pile-up suppression, particle ID and clustering are essential for maximising the physics performance of the Phase-II Global trigger in ATLAS. To address this, we train both convolutional and DeepSets neural networks to exploit cluster topologies to accurately predict calorimeter cell labels, and benchmark performance against existing approaches. We optimise the networks for firmware deployment and obtain resource and timing estimates.

Speakers: Alex Martynwood (UCL), Naoki Kimura (UCL), Nikos Konstantinidis (UCL), Noah Clarke Hall (University College London)

smarthep_poster.pdf

smarthep_talk.pdf

16:49 NextGen Trigger WP 2.2: Enhancing the L0 Muon Trigger: Project goals and needs

The NGT WP 2.2 aims to improve the robustness of the L0 muon trigger system and include additional trigger strategies for non-pointing signatures from decay of long-lived exotic particles implementing novel trigger strategies in firmware. Here we present the goals of the project, as well as the needs and requirements from available tools such as HLS4ML.

Speakers: Maria Carnesale (Sapienza Universita e INFN, Roma I (IT)), Oliver Kortner (Max Planck Society (DE)), Rimsky Alejandro Rojas Caballero (University of Massachusetts (US)), Verena Ingrid Martinez Outschoorn (University of Massachusetts (US))

SMARTHEP_poster.pdf

SMARTHEP_talk.pdf

16:52 Nanosecond AI for anomaly detection with decision trees on FPGA

We present an interpretable implementation of the autoencoding algorithm, used as an anomaly detector, built with a forest of deep decision trees on FPGA, field programmable gate arrays. Scenarios at the Large Hadron Collider are considered for which the autoencoder is trained using the Standard Model. The design is then deployed for anomaly detection of unknown processes. The inference is made with a latency value of 30 ns at percent-level resource usage using the Xilinx Virtex UltraScale+ VU9P FPGA. The work is documented at https://arxiv.org/abs/2304.03836

Speaker: Joerg Stelzer (University of Pittsburgh (US))

AD with fwX

16:55 Neural Architectures and Data Processing Pipelines for Irradiation Experiments: from the Automatic Assessment of Proposals to the Monitoring of the Beam Quality

Irradiation facilities, infrastructures for assessing devices and material radiation-hardness, face a variety of challenges, from the management of the experiment-selection process to the monitoring of the beam quality they need. While addressing vastly different issues, the answers may be found in carefully engineered Machine Learning and Artificial Intelligence (AI) solutions.
The applications of AI models in High-Energy-Physics (HEP) data analysis are well-established, in particular with neural networks and deep-learning algorithms. We suggest that recent advances in Natural-Language-Processing techniques such as transformer architectures may be used for the experiments’ proposals assessment and the development of new attention-based monitoring and anomaly detection tools used during their execution.
We provide supporting evidence for our approach by describing 1) how we help assess HEP-related scientific proposals within the RADNEXT EU-project and 2) how we monitor and evaluate the transverse beam profile quality in real-time at the CERN IRRAD facility in the EURO-LABS EU-project.

Speaker: Jaroslaw Szumega (CERN EP-DT-DD, Mines ParisTech (FR))

smarthep_edge_jszumega_flashtalk.pdf

smarthep_edge_jszumega_poster.pdf

17:00 Real-time search for Dark Photons at the Upgraded LHCb experiment (Poster Upload)

This work presents a new search for soft dark photons from charm decays, made possible by the novel real-time analysis (RTA) capabilities of the upgraded LHCb detector. The challenge consists in finding a peak on top of an irreducible non-resonant background of several kHz. In LHC Run 3, LHCb can read out the entire detector in real time (at 30 MHz) and filter interesting events through a two-stage software trigger using farms of GPUs (first stage) and CPUs (second stage). ML-based classification algorithms are employed at both stages to select charm decays, identify the extremely soft electrons that dark photons decay into, and reduce the overwhelming combinatorial background. The data throughput is further reduced by writing to disk only the interesting part of each event.

Speakers: Carlos Eduardo Cocha Toapaxi (Ruprecht Karls Universitaet Heidelberg), Martino Borsato (Universita & INFN, Milano-Bicocca (IT))

Poster_ML_CC.pdf

17:05 Flavour Tagging in Run 3 at the LHCb experiment (Poster Upload)

One of the main goals of the LHCb experiment is to study charge-parity violation by looking at the decays of the large variety of beauty mesons created in pp collisions at LHC. Such studies are particularly challenging in the presence of $B$$-\overline{B}$ oscillations as the $B$ meson flavour at production time might be different from the flavour at its decay time.

Flavour Tagging algorithms exploit the correlations between the $B$ meson production flavour and features of the global event to tag the candidate as $B$ or $\overline{B}$. Together with the tagging decision, the probability of a wrong tagging decision must be provided which is estimated through the application of Machine Learning algorithms.

The purpose of this contribution is to introduce the strategy and developments of the Flavour Tagging algorithms for Run 3.

Speaker: Micol Olocco (Technische Universitaet Dortmund (DE))

poster_FT.pdf

17:10 Track Reconstruction with Graph Neural Networks on Heterogeneous Architectures (Poster Upload)

The next decade will see an order of magnitude increase in data collected by high-energy physics experiments, driven by the High-Luminosity LHC (HL-LHC). The reconstruction of charged particle trajectories (tracks) has always been a critical part of offline data processing pipelines. The complexity of HL-LHC data will however increasingly mandate track finding in all stages of an experiment's real-time processing. This paper presents a GNN-based track-finding pipeline tailored for the Run 3 LHCb experiment's vertex detector and benchmarks its physics performance and computational cost against existing classical algorithms on GPU architectures. A novelty of our work compared to existing GNN tracking pipelines is batched execution, in which the GPU evaluates the pipeline on hundreds of events in parallel. We evaluate the impact of neural-network quantisation on physics and computational performance, and comment on the outlook for GNN tracking algorithms for other parts of the LHCb track-finding pipeline.

Speaker: Fotis Giasemis (Centre National de la Recherche Scientifique (FR))

giasemis_poster_edge_ml.pdf

17:12 Ultrafast Jet Classification at the HL-LHC (Poster Upload)

Three machine learning models are used to perform jet origin classification.
These models are optimized for deployment on a field-programmable gate array device.
In this context, we demonstrate how latency and resource consumption scale with the input size and choice of algorithm.
Moreover, the models proposed here are designed to work on the type of data and under the foreseen conditions at the CERN LHC during its high-luminosity phase.
Through quantization-aware training and efficient synthetization for a specific field programmable gate array, we show that $\mathcal{O}(100)$ ns inference of complex architectures such as Deep Sets and Interaction Networks is feasible at a relatively low computational resource cost.

Speaker: Denis-Patrick Odagiu (ETH Zurich (CH))

Ultrafast Jet Classification - EDGE ML School CERN.pdf

17:13 Optimizing Multi-Model Compression for Resource-Constrained Devices (Poster Upload)

We address the challenge of compressing a sequence of models for deployment on computing- and memory-constrained devices. This task differs from single model compression, as the decision to apply compression schemes either independently or jointly across all sub-networks introduces a new degree of freedom. We evaluate the performance of pruning and quantization techniques for model compression in the context of a prototypical image restoration and object detection multi-model system. We propose an adaptation of Quantization Aware Training (QAT) and pruning techniques, where the multi-model system is fine-tuned as a single unit with an adapted loss function, rather than applying these techniques to each model individually.

Speaker: João Luís Prado

prado_compression_edge.pdf

17:14 Licensing, Copyright and AI Training Data: Evidence from 2 Billion Images (Poster Upload)

We explore the implications of copyright on the composition, quality, and, therefore, inherent bias of AI training datasets. We study the example of LAION5B, a dataset of 5 billion images from the web, which is widely used in research and commercial applications of ML, including generative AI. We document the extent to which images in this dataset have a clear license status and to which extent it may be possible to infer the rightsholder. With this, we can compare the characteristics of a subset of images with either permissive licenses and/or identifiable rightsholders to the full dataset and estimate how the distribution of images and depicted concepts would change if copyright were enforced at scale. This allows us to contribute much-needed empirical evidence to a discussion of the role of copyright in enabling or restricting the availability of unbiased training datasets.

Speakers: Prof. Christian Peukert (University of Lausanne, HEC), Mr Irvin Dalaud (University of Lausanne, HEC), Viktor Shcherbakov (University of Lausanne, HEC)

Tuesday 24 September

09:00 → 11:00 Lectures: FPGAs

Convener: Sioni Paris Summers (CERN)

09:00 Introduction to FPGAs and FPGA inference

Speaker: Mario Ruiz Noguera (AMD)

cern_edge_ml_fpga_ai.pdf

09:45 AMD Brevitas QAT and the FINN compiler

Speaker: Mario Ruiz Noguera (AMD)

11:00 → 11:30 Coffee break

11:30 → 12:30 Lectures: Heterogeneous Accelerated Compute Cluster - HACC

Convener: Dylan Sheldon Rankin (University of Pennsylvania (US))

11:30 Heterogeneous Accelerated Compute Cluster

Speaker: Prof. Gustavo Alonso (ETHZ)

CERN-GA-ETH.pdf

CERN-GA-ETH.pptx

HACC

12:15 Discussion

12:30 → 13:45 Lunch break

13:45 → 16:00 Tutorial: FPGAs

Convener: Anna Sfyrla (Universite de Geneve (CH))

13:45 QONNX

Speakers: AMD/XILINX team, Marius Köppel (ETH Zurich (CH))

QONNX.pdf

14:15 Discussion

14:30 Tutorial: Working with QONNX

Speaker: Marius Köppel (ETH Zurich (CH))

15:15 Tutorial: Training a fully heterogeneously quantized NN with HGQ

Speaker: Chang Sun (California Institute of Technology (US))

HGQ-EdgeML.pdf

16:00 → 16:30 Coffee break

16:30 → 18:00 Tutorial: fwXmachina

Convener: Tae Min Hong (University of Pittsburgh (US))

16:30 Introduction to fwX

Speaker: Tae Min Hong (University of Pittsburgh (US))

TMHong20240920_SMARTHep_tutorial.pdf

17:00 Tutorial

Speakers: Ben Carlson, Elangovan Yuvaraj, Isabelle Eleonore Taylor (University of Pittsburgh (US)), Joerg Stelzer (University of Pittsburgh (US)), Pavel Serhiayenka, Rajat Gupta (University of Pittsburgh (US)), Santiago Cane (University of Pittsburgh (US)), Tae Min Hong (University of Pittsburgh (US))

Wednesday 25 September

08:45 → 09:30 Lectures: Generative AI in Research and Science

Convener: Andreas Salzburger (CERN)

08:45 Generative AI in Research and Science

The intersection of generative AI and science is revolutionizing research and discovery across various scientific disciplines. Generative AI, capable of creating new content based on learned patterns, accelerates hypothesis generation, drug discovery, material design, and personalized medicine by simulating experiments and predicting outcomes. It enhances climate science, environmental modeling, and scientific communication through advanced data analysis and visualization. However, this integration also presents challenges, including data quality, interpretability, and ethical concerns, necessitating careful management to ensure responsible and effective use of AI in scientific research.

We are exploring use cases and relevant examples of how GenAI has been adopted by the research community and how NVIDIA has invested in tools, libraries and models to enable open research and collaboration.

Speaker: Geetika Gupta (NVIDIA, Director Product Management)

GenAI for Science Examples-v1.pdf

09:15 Discussion

09:30 → 11:15 Tutorial: FPGAs

Convener: Sioni Paris Summers (CERN)

09:30 Training a quantized/pruned NN with QKeras/Tensorflow and deployment with hls4ml (via QONNX)[hls4ml team]

Speakers: Benjamin Ramhorst (ETHZ), Benjamin Ramhorst (Imperial College (GB))

hls4ml-tutorial-BR-SmartHEP.pdf

Join GitHub Team

Registration Form

Tutorial platform

11:15 → 11:45 Coffee break

11:45 → 12:45 Lectures: Edge ML out of HEP

Convener: Sioni Paris Summers (CERN)

11:45 Efficient architectures with geometric deep learning

Speaker: Hampus Otto Linander (Verses AI)

GDL_EdgeML_HampusLinander2.pdf

12:30 Discussion

12:45 → 14:00 Lunch break

14:00 → 15:00 Lectures: Neuromorphic computing

Convener: Dr Jean-Roch Vlimant (California Institute of Technology (US))

14:00 Neuromorphic computing

Speakers: Stanislaw Wozniak (IBM Research), Thomas Ortner (IBM Research)

Lecture.pdf

14:45 Discussion

15:00 → 15:30 Coffee break

15:30 → 17:00 Tutorial: Neuromorphic computing

Convener: Dr Jean-Roch Vlimant (California Institute of Technology (US))

15:30 Tutorial on neuromorphic computing

Speaker: IBM Research

Tutorial link

17:00 → 18:00 Lectures: Satellite

Convener: Dr Jean-Roch Vlimant (California Institute of Technology (US))

17:00 Real time applications in earth-monitoring satellites

Speakers: Agenium Space, François Devieilleville (Agenium)

devieilleville_smartHEP.pdf

17:30 Discussion

Thursday 26 September

09:00 → 10:00 Tutorial: FPGAs

Convener: Martino Borsato (Universita & INFN, Milano-Bicocca (IT))

09:00 Decision Forests on FPGAs with conifer

Speaker: Sioni Paris Summers (CERN)

conifer at gitlab

sps_conifer_smarthep.pdf

09:45 Discussion

10:00 → 11:00 Lectures: Greenwaves Technologies

Convener: Martino Borsato (Universita & INFN, Milano-Bicocca (IT))

10:00 Greanwaves Technologies - ultralow power AI processors

Speaker: Marco Fariselli (Greenwave technologies)

greenwaves_smarthep2024.pdf

Greenwaves technologies

10:45 Discussion

11:00 → 11:30 Coffee break

11:30 → 12:30 Lectures: GROQ

Convener: Sioni Paris Summers (CERN)

11:30 GROQ: Inference at the speed of data

Speaker: Tobias Becker (Maxeler Technologies)

Groq is Fast AI Inference

12:30 → 13:45 Lunch break

13:45 → 15:10 Lectures

Convener: Maurizio Pierini (CERN)

13:45 Real-time processing for gravitational wave physics

Speaker: Eric Anton Moreno (Massachusetts Institute of Technology (US))

EdgeML_ML4GW_EMoreno.pdf

14:15 Discussion

14:25 Real-time processing at the LST of CTAO

Speaker: Matthieu Heller (Universite de Geneve (CH))

CTAO_RealTime_Heller_20240925.pdf

14:55 Discussion

15:10 → 16:00 Closing session

Conveners: Anna Sfyrla (Universite de Geneve (CH)), Maurizio Pierini (CERN), Sioni Paris Summers (CERN), Thea Aarrestad (ETH Zurich (CH))

15:15 Poster prizes

PosterAwards_slides.pdf

15:30 Closing talk

ClosingSlides.pdf

15:54 Further references

a3d3

Fast ML conference 2024

Fast ML lab

NextGen 1st technical workshop

NextGen triggers

SMARTHEP ITN

Friday 27 September

09:00 → 18:00 NGT workshop: hls4ml HEP community forum

hls4ml HEP Community Forum